Joint Theory Seminar: Romuald Janik

Speaker: Romuald Janik

Title: From Machine Learning to Physics, and back again...  

Abstract: In this talk I would like to describe some fruitful interrelations between Machine Learning and Physics.
On the one hand, I will show how to use the tools of machine learning to estimate the entropy (and free energy) of a system directly from Monte Carlo configurations at a given temperature, which is commonly believed to be extremely difficult if not impossible by conventional means.
On the other hand, I will describe a proposed definition of complexity for deep neural networks, which is based on some intuitions from physics. I will show how one can use it together with a complementary notion of effective dimension to quantify the intuitive difficulty of a dataset or a learning task. These notions also reveal a rather mysterious power-law scaling during training.